GlossaryUpdated
Intelligence explosion
An intelligence explosion is the idea that an AI able to improve itself could set off a fast chain of ever-smarter versions, leaving human intelligence far behind in a short time.
An intelligence explosion is the idea that an AI able to improve itself could set off a fast chain of ever-smarter versions: each version builds a better one, which builds a better one still, until AI is far ahead of human intelligence. The word "explosion" is about speed. The claim is not just that AI gets smarter, which it already does, but that the improvement could feed on itself and arrive faster than people can react.
No intelligence explosion has happened. It is a prediction, and one of the most argued-about ones in AI. You will see it in headlines about AI companies, in the debate about recursive self-improvement, and in warnings about artificial superintelligence.
What it means for you at work
For your work this week, nothing changes. The AI tools you use are not improving themselves, and they still make mistakes a careful colleague would catch.
What the idea does change is how to read the news. Company announcements, investor documents and opinion pieces now use phrases like "intelligence explosion", "takeoff" and "self-improving AI" more often, and not always with the same meaning. Knowing the term lets you ask the useful question: is this a claim that AI helps build AI (true today), or a claim that the loop is speeding itself up without people (not shown)?
Where the term comes from
In 1965 the British statistician I. J. Good, who had worked with Alan Turing on code-breaking during the Second World War, wrote an essay called "Speculations Concerning the First Ultraintelligent Machine". His argument fits in three steps:
- Imagine a machine that beats the cleverest person at every thinking task.
- Designing machines is a thinking task, so this machine could design a better machine than people can.
- That better machine could design a better one still, and there would "unquestionably be an intelligence explosion".
He then added the line people still quote: the first ultraintelligent machine "is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control". The worry about control was there from the first page.
In 1993 the science fiction writer and computer scientist Vernor Vinge named the result the "technological singularity": a point after which the future cannot be predicted from the past. In 2014 the philosopher Nick Bostrom's book Superintelligence set out the modern version of the argument and gave it its vocabulary of takeoff speeds.
Seed AI: the first step of the chain
A seed AI is the version that starts the chain. It does not need to be superintelligent. It only needs to be good enough at improving AI that its successor is better at the same job. The AI writer Eliezer Yudkowsky used the term in his early 2000s writing about how a self-improving AI might be built.
The idea explains why people watch AI's ability to do AI research so closely. A model that writes better essays does not start a chain. A model that makes the next model better at improving models might. That is what RSI evals try to measure, and why tests of AI research skill get so much attention. See LLM evals for how such tests work and where they mislead.
No AI today is a seed AI in the strong sense. People still choose what to try, approve the changes and run the training.
Fast takeoff vs slow takeoff
Most of the argument about an intelligence explosion is about speed, not about whether AI keeps improving. Bostrom described three speeds:
- Slow: the change from human level to far beyond it takes decades or centuries. Society has time to adjust, argue and pass laws.
- Moderate: it takes months or years. Some groups can react; many cannot.
- Fast: it takes minutes, hours or days. Nobody has time to react once it starts.
A fast takeoff is also called a hard takeoff, and a slow one a soft takeoff. In 2008 Yudkowsky and the economist Robin Hanson argued the question in a long public exchange, later published as "The Hanson-Yudkowsky AI-Foom Debate", with Yudkowsky expecting a sudden jump and Hanson expecting gradual change spread across many companies.
In 2018 the AI researcher Paul Christiano put the slow view in a testable form. He predicted that before AI could double the world economy in a single year, there would first be a full four-year period in which the world economy doubled. In other words, there would be visible warning signs first.
The case for and against
Arguments that the chain could run fast:
- Software copies instantly. A better AI researcher, once it exists, can be run as thousands of copies.
- AI already does real AI work. Companies say their models write a large share of their own code and help run experiments.
- Each gain helps the next. That is the definition of the loop, and small gains repeated many times add up.
Arguments that it will be slow or stall:
- Experiments take time. Every new idea still has to be tested by training a model on expensive hardware, and smarter software does not make chips arrive faster.
- Some knowledge needs the real world. Data about people, markets and physics comes from watching the world, which cannot be sped up much.
- The easy gains come first. Each later improvement is harder to find, which can slow the chain however clever the AI is.
- Testing is hard. A loop only keeps a change if the new version scores higher. A flawed test lets the AI improve at the test instead of at the work.
What AI companies said in 2026
The idea moved from essays into company documents this year. In June 2026 Anthropic published a report on recursive self-improvement built on its own engineering data. On 2 November 2026, in an updated S-1 filing (the document a US company files before selling shares to the public), Anthropic wrote that its estimates for future models indicate it "may imminently meet" its own policy's definition of "dramatically accelerated progress attributable to AI". In the same paragraph it said no current model has crossed that threshold, and that it does not expect to let Claude build new frontier models without "meaningful human supervision or understanding". The RSI page covers the filing in more detail.
Note what the filing does and does not say. It describes faster progress with AI's help, and a threshold the company expects to reach. It does not claim an intelligence explosion has started.
How to read an intelligence explosion headline
When an article uses the term, four questions sort the claim:
- Help or loop? Is the claim that AI helps people build AI, or that AI improves itself in a chain? The first is true today. The second has not been shown.
- Who decides? Are people still choosing and approving the changes? If yes, it is not the loop in Good's sense.
- What is measured? A named test, a number from the company, or an opinion?
- How fast? Does it give a timescale, and is it slow, moderate or fast in Bostrom's terms?
A prompt that applies the same check to anything you paste in:
You are a careful editor.
<text>[paste the headline, paragraph or announcement]</text>
1. In one sentence, what is claimed: AI helping people build AI, or AI improving itself in a loop with people out of the decisions?
2. Who approves the changes in what is described: people, the AI, or not stated?
3. What evidence is given: a named test with a number, a company statement, an example, or none?
4. Is a timescale given? Classify it as slow (decades), moderate (months to years), fast (days) or none.
Quote the exact words you are judging. Do not add facts that are not in the text.
The last line keeps the check honest. Without it, the AI tends to fill gaps with what it believes about the company or the field.
Questions
What is an intelligence explosion in AI?
It is the idea that once an AI is good enough to improve its own design, each improved version will be better at improving the next one. If that chain speeds up, AI could go from roughly human level to far beyond it in a short time. It is a prediction about what might happen, not something that has happened.
Who came up with the term intelligence explosion?
The British statistician I. J. Good, in a 1965 essay called Speculations Concerning the First Ultraintelligent Machine. He wrote that such a machine could design even better machines, and that there would then 'unquestionably be an intelligence explosion'.
What is a seed AI?
A seed AI is the first AI that is good enough to start improving itself: the version that builds version two, which then builds version three. The term was used by the AI writer Eliezer Yudkowsky around 2001. No AI today is a seed AI in this sense, because people still choose the experiments, approve the changes and run the training.
Is an intelligence explosion the same as the singularity?
They are close but not the same. The intelligence explosion is the mechanism: AI improving AI in a fast chain. The technological singularity, a term made popular by the writer Vernor Vinge in 1993, is the claimed result: a point after which the future becomes impossible to predict.
What is the difference between a fast and a slow takeoff?
A fast, or hard, takeoff means AI jumps from about human level to far beyond it in days or months. A slow, or soft, takeoff means the same change happens over years or decades, with time for people and institutions to react. Most of the argument about an intelligence explosion is about which of these is more likely.
Has the intelligence explosion started?
No one has shown it. AI already helps build AI: it writes code and runs experiments at AI companies. But the defining feature, a loop that speeds itself up without people in charge, has not been seen. In November 2026 Anthropic said in a stock market filing that it may soon meet its own threshold for AI-accelerated progress, and that no current model has crossed it.